Bottleneck Identification in Semiconductor Manufacturing: A Machine Learning Approach

Abstract

The semiconductor industry is known for its complex production. Thousands of machines (tools) perform thousands of operations over a diverse range of products with re-entrant flows and shifting bottlenecks. Detecting and predicting these bottlenecks in real-time and promptly responding to them at the shop-floor level can make a big difference in terms of efficiently utilizing the high capital equipment and achieving operational excellence. So far, approaches to this problem have used traditional optimization methods. This paper pioneers the application of big data ML models to the problem, tested on industry data. We propose a dual-phased approach that leverages Machine Learning (ML) for this task. The first phase involves evaluating the most relevant production parameters (features) in the production line, along with their predefined target values, while the second phase enhances and refines these parameters at the segment level of the production line to reflect aspects of balancing and leveling. We show that the most effective ML model is attained using the XGBoost algorithm, which achieves 95% accuracy. This outcome enables a more precise classification of the dynamically shifting bottlenecks within the production line relative to other methods in practice, thereby allowing production staff better control and performance in the long run.